Grok users hit word-salad glitch

This looks like a reliability regression, not evidence of a new model launch—but opaque rollout experiments make the Grok 4.6 theory plausible.

• –Gibberish output is a serious trust problem for a general-purpose assistant, even when recovery only requires refreshing a session
• –The apparent limitation to some Grok Lite users suggests a model-routing, decoding, or serving-path issue rather than a universal model failure
• –Grok 4.6 launched through developer surfaces on August 12, but no public evidence connects it to this incident
• –Developers using Grok in production should add output validation, retry logic, and anomaly detection for low-entropy or incoherent generations